Uploaded November 2026 | Updated November 2026, 3 days ago
Connect with the cloud native community and explore CNCF events: kubecon.io
The Storage Bottleneck AI Didn't See Coming: Inside StepFun’s Multimodal Production Stack - Joe Zhou, Juicedata Inc. & Frank Li, StepFun
Multimodal AI has outgrown the storage systems built for it. Training and inference now mix small-file storms, high-concurrency random reads, terabyte-scale checkpoints, and 100 GB+ model files across heterogeneous GPU clusters, private data centers, and public clouds. The hard part isn't any one workload. It means serving all of them through a single storage access layer that follows compute wherever it runs, whether on-prem or in any cloud. In this keynote, we'll show how StepFun made that move on Kubernetes, TiKV, and JuiceFS. Hardware-aware CSI mounts handle heterogeneous nodes, TiKV-backed metadata scales to more than 10,000 mounting hosts, point-get optimization speeds up relevant read-only operations by roughly 50%, and tuning both the metadata engine and readahead for Safetensors loads cuts read amplification by roughly 71%. We'll share how these improvements were merged upstream to turn production lessons into community gains.
Connect with the cloud native community and explore CNCF events: kubecon.io
The Storage Bottleneck AI Didn't See Coming: Inside StepFun’s Multimodal Production Stack - Joe Zhou, Juicedata Inc. & Frank Li, StepFun
Multimodal AI has outgrown the storage systems built for it. Training and inference now mix small-file storms, high-concurrency random reads, terabyte-scale checkpoints, and 100 GB+ model files across heterogeneous GPU clusters, private data centers, and public clouds. The hard part isn't any one workload. It means serving all of them through a single storage access layer that follows compute wherever it runs, whether on-prem or in any cloud. In this keynote, we'll show how StepFun made that move on Kubernetes, TiKV, and JuiceFS. Hardware-aware CSI mounts handle heterogeneous nodes, TiKV-backed metadata scales to more than 10,000 mounting hosts, point-get optimization speeds up relevant read-only operations by roughly 50%, and tuning both the metadata engine and readahead for Safetensors loads cuts read amplification by roughly 71%. We'll share how these improvements were merged upstream to turn production lessons into community gains.










